Training

The script train.py is all you need :)

Important Command Line Arguments for train.py

--source_path / -s

Path to the source directory containing the dataset.

--model_path / -m

Path where the trained model should be stored.

--iterations

The total number of iterations for training (30_000 by default).

--warm_up

The iteration index until which MLP optimization is paused (3000 by default).

--warm_up_3d_features

The iteration index until which the Gaussian features start to optimized. (10000 by default) Need to use together with the iterative_opt_interval.

--iterative_opt_interval

The default mode is optimizaing only the colors and geometry of the scene. After iterative_opt_interval iteration, the mode changes to optimizing only the Gaussian features.

--monitor_mem

Configure to monitor the RAM and CUDA used.

--lambda_reg_deform

Apply regularization to the deformation of the Gaussians. (0.0 by default)

--num_sampled_pixels

Number of sampled pixels per image for contrastive semantically-aware learning. (5000 by default)

--num_sampled_masks

Number of sampled masks per image for contrastive semantically-aware learning. (50 by default)

--smooth_K

Number of neighbors for computing smooth Gaussian features (16 by default)

--load2gpu_on_the_fly

Configure to load images / masks to VRAM on the fly. (For training on small GPU)

--load_image_on_the_fly

Configure to load images from storage to RAM on the fly (For training on limited RAM)

--load_mask_on_the_fly

Configure to load masks from storage to RAM on the fly (For training on limited RAM)

--eval

Add this flag to do training/test split for evaluation.

--white_background / -w

Add this flag to use white background instead of black (default)

--test_iterations

Space-separated iterations at which the training script computes L1 and PSNR over test set.

--save_iterations

Space-separated iterations at which the training script saves the Gaussian model.

--densify_until_iter

Iteration where densification stops, 15_000 by default.

--contrastive_mode

Select different mode for contrastive learning. Default: soft (available modes: soft, all, hard)

--mask_black_bg

Configure to mask out loss on black background pixels. Only used in Technicolor dataset.

NeRF-DS

python train.py -s data/NeRF-DS/<NAME> -m output/NeRF-DS/<NAME> --warm_up 3000 --warm_up_3d_features 15000 --iterative_opt_interval 20000 --iterations 30000 --test_iterations 5000 10000 20000 30000 --save_iterations 20000 30000 --monitor_mem --densify_until_iter 15000 --lambda_reg_deform 0.0 --eval --num_sampled_pixels 5000 --num_sampled_masks 25 --smooth_K 16 --contrastive_mode soft ## You can configure --load2gpu_on_the_fly --load_image_on_the_fly --load_mask_on_the_fly for running on smaller GPU or local machine having less RAM.

HyperNeRF

python train.py -s data/HyperNeRF/<interp/misc>/<NAME> -m output/HyperNeRF/<NAME> --warm_up 1500 --warm_up_3d_features 15000 --iterative_opt_interval 20000 --iterations 30000 --test_iterations 5000 10000 15000 20000 30000 --save_iterations 20000 30000 --monitor_mem --densify_until_iter 9000 --lambda_reg_deform 0.0 --eval --num_sampled_pixels 5000 --num_sampled_masks 25 --smooth_K 16 --contrastive_mode soft ## You can configure --load2gpu_on_the_fly --load_image_on_the_fly --load_mask_on_the_fly for running on smaller GPU or local machine having less RAM.

Neu3D

python train.py    -s data/Neu3D/<NAME> -m output/Neu3D/<NAME> --warm_up 3000 --warm_up_3d_features 15000 --iterative_opt_interval 20000 --iterations 30000 --test_iterations 10000 15000 20000 30000 --save_iterations 10000 15000 20000 30000 --monitor_mem --densify_until_iter 8000 --lambda_reg_deform 0 --eval --load2gpu_on_the_fly --num_sampled_pixels 10000 --num_sampled_masks 50 --smooth_K 16 --contrastive_mode soft --load_mask_on_the_fly --load_image_on_the_fly ## For multiview dataset, it's suggested to load images and anything-masks on-the-fly to reduce RAM usage

Immersive

python train.py    -s data/immersive/<NAME> -m output/immersive/<NAME> --warm_up 1000 --warm_up_3d_features 15000 --iterative_opt_interval 20000 --iterations 30000 --test_iterations 5000 10000 15000 20000 30000 --save_iterations 10000 15000 20000 30000 --monitor_mem --densify_until_iter 3000 --lambda_reg_deform 0 --eval --load2gpu_on_the_fly --num_sampled_pixels 10000 --num_sampled_masks 50 --contrastive_mode soft --load_mask_on_the_fly --load_image_on_the_fly --end_frame 50 ## For multiview dataset, it's supported to load images and anything-masks on-the-fly to reduce RAM usage

Technicolor

python train.py    -s data/technicolor/Undistorted/<NAME> -m output/technicolor/<NAME> --warm_up 3000 --warm_up_3d_features 15000 --iterative_opt_interval 20000 --iterations 30000 --test_iterations 5000 10000 15000 20000 30000 --save_iterations 10000 15000 20000 30000 --monitor_mem --densify_until_iter 5000 --lambda_reg_deform 0 --eval --load2gpu_on_the_fly --num_sampled_pixels 10000 --num_sampled_masks 50 --contrastive_mode soft --load_mask_on_the_fly --load_image_on_the_fly --mask_black_bg ## For multiview dataset, it's supported to load images and anything-masks on-the-fly to reduce RAM usage